Anthropic blocks possible attempt to use AI to make biological weapons

Anthropic Says It Stopped AI Misuse Linked to Biological Weapons Risks

Ninoda.com – Anthropic says it has interrupted several efforts to use its Claude artificial intelligence system in ways that could contribute to the creation of biological weapons, highlighting a threat the company describes as among the gravest dangers posed by advanced AI.

The cases appear in Anthropic’s latest threat intelligence report, released on Thursday. The document covers activity identified between December 2025 and August 2026, ranging from scams and surveillance operations to cyber espionage, influence campaigns and potential weapons-related research.

Five case studies involved people using Claude in ways that could assist biological weapons development. Anthropic said this area demands particular caution because frontier AI systems can make highly technical information easier to access and organize.

The same information that can be used to develop a biological weapon could also be used to develop, for example, a vaccine or a cure for a disease.

That overlap creates a difficult challenge for AI companies. Scientific knowledge can have legitimate medical, academic and public-health applications while also carrying potential for harmful misuse. Anthropic said it had blocked researchers whose use of its tools crossed into activity that could support biological weapons work.

A widening range of AI misuse

The report describes alleged misuse of Claude Haiku, Sonnet and Opus models. Anthropic said none of the identified incidents involved Claude Fable or its powerful Mythos-class models, except for one distillation case. Distillation is a technique in which a smaller, less costly model is trained with help from a larger model.

Anthropic said the detected activity was not limited to biological risks. The company listed fake dating applications, hotel Wi-Fi fraud, systems designed to identify dissidents, criminal schemes, spyware activity and politically motivated operations among the cases it disrupted.

Its review also identified six instances in which Claude was used to develop software connected with conventional weapons. These included firearms, missiles, armed drones, bombs, other munitions, and systems used to target or control them.

Actors linked to a Russia-based cyber espionage effort and an Iranian propaganda institution also used Claude, Anthropic said. The company further accused Chinese AI firms of attempting to reproduce Claude’s capabilities.

Cybercrime and state-backed activity

AI tools are increasingly being folded into cyber operations because they can help users write, revise and analyze software at speed. Anthropic said cybercriminals and state-supported hacking groups had turned to its technology for assistance in their operations.

ShinyHunters was named in the report, along with China-based laboratories. Anthropic also described activity consistent with the Russia-based hacking group Midnight Blizzard. In that case, a group allegedly used AI to create a system that noticed when security tools had identified its malware, then repeatedly altered the code in an attempt to avoid detection.

The significance of such activity is not simply that AI can generate code. Automated rewriting and testing may allow malicious operators to cycle through changes more rapidly, placing additional pressure on defenders who must recognize and block new versions of harmful software.

Anthropic said it has used lessons from the cases to strengthen its ability to prevent, identify and disrupt future abuse. It also said intelligence was shared with relevant authorities and industry partners when appropriate.

Pressure grows over frontier AI safety

Threat intelligence reports have become a recurring feature of the AI industry as major developers seek to show how they monitor misuse of their systems. Google recently disclosed an attempted request involving Gemini for a detailed technical guide to synthesizing weaponized biological agents.

The new Anthropic report arrives amid growing arguments over whether the development of increasingly capable AI should be slowed. US Senator Bernie Sanders has called for a pause in advanced AI development and legislation banning artificial superintelligence.

President Donald Trump has taken a different position. Speaking on Thursday, he said he was concerned that failing to lead in AI would leave the United States at a serious disadvantage.

If we don’t win AI, we’re going to be put in a very bad position.

The debate is also shaped by warnings from within Anthropic itself. Earlier this year, one of the company’s leading safety researchers said the rapid pace of AI progress created a greater than 10% chance that the technology could kill all humans.

Anthropic’s report does not suggest that every difficult scientific request is malicious, nor does it treat all use of advanced AI as dangerous. Instead, it illustrates why developers face a complex task: supporting useful research and everyday productivity while preventing systems from lowering barriers for fraud, repression, cyberattacks or weapons development.

For readers and organizations using AI tools, the report is a reminder that these systems are no longer merely experimental writing assistants. Their growing ability to process technical material, produce software and support research makes safeguards, human oversight and rapid responses to misuse central questions for the technology’s future.